Process mapping and anomaly detection in laser wire directed energy deposition additive manufacturing using in-situ imaging and process-aware machine learning
Anis Assad, Benjamin Bevans, Willem Potter, Prahalada Rao, Denis Cormier, Fernando Deschamps, Jakob D. Hamilton, Iris V. Rivero
Abstract
• Directed energy deposition meltpool data was collected from 128 unique process parameter combinations using in-situ, high-speed imaging. • Six physically intuitive meltpool features based in meltpool morphology and intensity were extracted from in-situ images. • Machine learning models were trained to classify meltpool features into one of four possible regimes: stable melting, stubbing, dripping, or incomplete melting. • Simple machine learning models achieved nearly 90% statistical fidelity, comparable to complex and computationally intense deep learning algorithms. This work concerns the laser wire directed energy deposition (LW-DED) additive manufacturing process. The objectives were two-fold: (1) process mapping – demarcating the process states as a function of the processing parameters; and (2) process monitoring – detecting process anomalies (instabilities) using data acquired from an in-situ meltpool imaging sensor. The LW-DED process enables high-throughput, near-net shape manufacturing. Without rigorous parameter control, however, LW-DED often introduces defects due to stochastic process drifts. To enhance scalability and reliability, it is essential to understand how LW-DED parameters affect processing regimes, and detect deleterious process drifts. In this work, single-track experiments were conducted over 128 combinations of laser power, scanning velocity, and linear mass density. Four process states were observed via high-speed imaging and delineated as stable, dripping, stubbing, and incomplete melting regimes. Physically intuitive meltpool features were used to train simple machine learning models for classifying the process state into one of the four regimes. The approach was benchmarked against computationally intense, black-box deep machine learning models that directly use as-received meltpool images. Using only six intuitive meltpool morphology and intensity signatures, the approach classified the LW-DED process state with statistical fidelity approaching 90 % (F1-score) compared to F1-score 87 % for deep learning models.